-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmain.py
executable file
·308 lines (254 loc) · 17.2 KB
/
main.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
import tensorflow as tf
import numpy as np
import scipy.misc
import random
from random import shuffle, randint
import sys
import os
import random
import gzip
import time
import pickle
import argparse
import models
from utils.ops import Ops
from models.model import Model
os.environ["CUDA_VISIBLE_DEVICES"]="1"
class CDRD:
def __init__(self,sess,args):
self.start_time = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
self.sess = sess
if not os.path.exists(args.sample_path):
os.makedirs(args.sample_path)
os.makedirs(args.sample_path+'Z/')
os.makedirs(args.sample_path+'dm1/')
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
self.G_model_filepath = args.model_path + 'G.ckpt'
self.D_model_filepath = args.model_path + 'D.ckpt'
if args.status=='test' :
if not os.path.exists(args.sample_path+args.restore_num+'/'):
os.makedirs(args.sample_path+args.restore_num+'/')
os.makedirs(args.sample_path+'Z/')
os.makedirs(args.sample_path+'dm1/')
self.G_model_filepath = self.G_model_filepath + args.restore_num
self.D_model_filepath = self.D_model_filepath + args.restore_num
self.ops = Ops()
self.model = Model(sess,args.summaries_path, args.batch_size, args.input_z_size,args.emb_size, args.class_num, args.code_dim, args.img_size)
self.train(args)
#============== Train ==================#
def train(self,args):
# Load Data
with gzip.open(args.data_dm1_path, 'rb') as f:
dm1_train_set, dm1_valid_set, dm1_test_set = pickle.load(f)
self.dm1_train, self.dm1_lb_train = dm1_train_set
self.dm1_test, self.dm1_lb_test = dm1_test_set
self.dm1_val, self.dm1_lb_val = dm1_valid_set
# MNIST choose 2000 ; USPS choose 1800
self.dm1_train = np.concatenate( (self.dm1_train, self.dm1_val) , axis=0)
self.dm1_lb_train = np.concatenate( (self.dm1_lb_train, self.dm1_lb_val), axis=0)
self.dm1_train_choose_idx = np.random.choice(len(self.dm1_train), 2000, replace=False)
self.dm1_test_choose_idx = np.random.choice(len(self.dm1_test), 2000, replace=False)
self.dm1_train = self.dm1_train[self.dm1_train_choose_idx]
self.dm1_test = self.dm1_test[self.dm1_test_choose_idx]
self.dm1_lb_train = self.dm1_lb_train[self.dm1_train_choose_idx]
self.dm1_lb_test = self.dm1_lb_test[self.dm1_test_choose_idx]
self.dm1_train = self.dm1_train.reshape((-1, 1,28,28))
self.dm1_test = self.dm1_test.reshape((-1, 1,28,28))
self.dm1_lb_train = self.ops.idx2one_hot(self.dm1_lb_train.shape[0], args.class_num, self.dm1_lb_train)
self.dm1_lb_test = self.ops.idx2one_hot(self.dm1_lb_test.shape[0],args.class_num, self.dm1_lb_test)
# For GAN dm1
self.dm1_2_train_choose_idx = np.random.choice(len(self.dm1_train), 2000, replace=False)
self.dm1_2_train = self.dm1_train[self.dm1_2_train_choose_idx]
self.dm1_2_train = self.dm1_2_train.reshape((-1, 1,28,28))
with gzip.open(args.data_dm2_path) as f:
dm2_train_set,dm2_test_set = pickle.load(f)
self.dm2_train, self.dm2_lb_train = dm2_train_set
self.dm2_test , self.dm2_lb_test = dm2_test_set
self.dm2_train_choose_idx = np.random.choice(len(self.dm2_train), 1800, replace=False)
self.dm2_test_choose_idx = np.random.choice(len(self.dm2_test), 1800, replace=False)
self.dm2_train = self.dm2_train[self.dm2_train_choose_idx]
self.dm2_test = self.dm2_test[self.dm2_test_choose_idx]
self.dm2_lb_train = self.dm2_lb_train[self.dm2_train_choose_idx]
self.dm2_lb_test = self.dm2_lb_test[self.dm2_test_choose_idx]
self.dm2_lb_train = self.ops.idx2one_hot(self.dm2_lb_train.shape[0],args.class_num, self.dm2_lb_train)
self.dm2_lb_test = self.ops.idx2one_hot(self.dm2_lb_test.shape[0],args.class_num, self.dm2_lb_test)
# For GAN dm2
self.dm2_2_train_choose_idx = np.random.choice(len(self.dm2_train), 1800, replace=False)
self.dm2_2_train = self.dm2_train[self.dm2_2_train_choose_idx]
self.dm2_2_train = self.dm2_2_train.reshape((-1, 1,28,28))
# Iteration number and data number
self.dm1_im_train_num = len(self.dm1_train)
self.dm1_im_train_iter_num = self.dm1_im_train_num / args.batch_size
self.dm1_2_im_train_num = len(self.dm1_2_train)
self.dm1_2_im_train_iter_num = self.dm1_2_im_train_num / args.batch_size
self.dm2_im_train_num = len(self.dm2_train)
self.dm2_im_train_iter_num = self.dm2_im_train_num / args.batch_size
self.dm2_2_im_train_num = len(self.dm2_2_train)
self.dm2_2_im_train_iter_num = self.dm2_2_im_train_num / args.batch_size
self.dm1_im_test_num = len(self.dm1_test)
self.dm1_im_test_iter_num = self.dm1_im_test_num / args.batch_size
self.dm2_im_test_num = len(self.dm2_test)
self.dm2_im_test_iter_num = self.dm2_im_test_num / args.batch_size
## Opt.
with tf.name_scope('Optimizer'):
with tf.name_scope('Discriminator'):
with tf.name_scope('Domain_1'):
self.dm1_d_gan_train_op = tf.train.AdamOptimizer(learning_rate=args.learning_rate,
beta1=0.5,
beta2=0.999,
epsilon=1e-08,
name='Adam1').minimize(self.model.D_loss+self.model.D_loss2 , var_list=self.model.d_d1_vars + self.model.d_d2_vars )
self.dm1_d_gan_train_op2 = tf.train.AdamOptimizer(learning_rate=args.learning_rate,
beta1=0.5,
beta2=0.999,
epsilon=1e-08,
name='Adam2').minimize(self.model.D_loss+self.model.D_loss2 , var_list= self.model.d_d1_vars + self.model.g_d1_vars + self.model.d_d2_vars + self.model.g_d2_vars )
with tf.name_scope('Generator'):
with tf.name_scope('Domain_1'):
self.dm1_g_train_op = tf.train.AdamOptimizer(learning_rate=args.learning_rate ,
beta1=0.5,
beta2=0.999,
epsilon=1e-08,
name='Adam3').minimize(self.model.G_loss+self.model.D_loss2 , var_list=self.model.g_d1_vars + self.model.g_d2_vars )
with tf.name_scope('Initial'):
init_op = tf.global_variables_initializer()
self.sess.run(init_op)
# restore the model(parameter )
self.G_saver = tf.train.Saver(var_list=self.model.g_vars, max_to_keep=2)
self.D_saver = tf.train.Saver(var_list=self.model.d_vars, max_to_keep=2)
print('Saver Build!')
if args.restore_mode == True:
self.G_saver.restore(self.sess, self.G_model_filepath)
print("G Model restored in file: %s" % self.G_model_filepath)
print('Restore G!')
self.D_saver.restore(self.sess, self.D_model_filepath)
print("D Model restored in file: %s" % self.D_model_filepath)
print('Restore D!')
# =========== Start Train and Val. =========== #
if args.status == 'train':
for i in range(0,args.epoch):
sample_z = self.ops.sample_Z (args.batch_size,args.input_z_size)
sample_z_lb,_ = self.ops.random_one_hot(args.class_num, args.batch_size)
self.dm1_1_img_train_idx = self.ops.shuffle_train_idx(self.dm1_im_train_num)
for dm1_k in range(0,self.dm1_im_train_iter_num):
stage = dm1_k+i*self.dm1_im_train_iter_num
self.dm1_2_im_train_iter_num = self.dm1_2_im_train_num / args.batch_size
self.dm2_im_train_iter_num = self.dm2_im_train_num / args.batch_size
self.dm2_2_im_train_iter_num = self.dm2_2_im_train_num / args.batch_size
dm2_k = dm1_k % self.dm2_im_train_iter_num
dm1_2_k = dm1_k % self.dm1_2_im_train_iter_num
dm2_2_k = dm1_k % self.dm2_2_im_train_iter_num
if dm2_k == 0 :
self.dm2_1_img_train_idx = self.ops.shuffle_train_idx(self.dm2_im_train_num)
if dm1_2_k == 0 :
self.dm1_2_img_train_idx = self.ops.shuffle_train_idx(self.dm1_2_im_train_num)
if dm2_2_k == 0 :
self.dm2_2_img_train_idx = self.ops.shuffle_train_idx(self.dm2_2_im_train_num)
# For disentanglement - dm1
batch_dm1_1_img = self.ops.pre_process(self.ops.get_batch(self.dm1_train , args.batch_size, self.dm1_1_img_train_idx , dm1_k ), if_random_flop= False).transpose(0,2,3,1)
batch_dm1_lb = (self.ops.get_batch(self.dm1_lb_train , args.batch_size, self.dm1_1_img_train_idx , dm1_k ))
# For disentanglement - dm2
batch_dm2_1_img = self.ops.pre_process(self.ops.get_batch(self.dm2_train , args.batch_size, self.dm2_1_img_train_idx , dm2_k ), if_random_flop= False).transpose(0,2,3,1)
batch_dm2_lb = (self.ops.get_batch(self.dm2_lb_train , args.batch_size, self.dm2_1_img_train_idx , dm2_k ))
# Update Encoder & Generator
for _ in range(args.update_time):
_,G_loss = self.sess.run([self.dm1_g_train_op,self.model.G_loss],
feed_dict = {self.model.X1 : batch_dm1_1_img,
self.model.X2 : batch_dm2_1_img,
self.model.X1_lb : batch_dm1_lb,
self.model.X2_lb : batch_dm2_lb,
self.model.z : sample_z,
self.model.z_lb : sample_z_lb })
_,D_gan_loss,D_dis_loss= self.sess.run([self.dm1_d_gan_train_op,self.model.D_loss,self.model.D_loss2],
feed_dict = {self.model.X1 : batch_dm1_1_img,
self.model.X2 : batch_dm2_1_img,
self.model.X1_lb : batch_dm1_lb,
self.model.X2_lb : batch_dm2_lb,
self.model.z : sample_z,
self.model.z_lb : sample_z_lb })
# Summary
train_sum = self.sess.run(self.model.train_sum,
feed_dict = {self.model.X1 : batch_dm1_1_img,
self.model.X2 : batch_dm2_1_img,
self.model.X1_lb : batch_dm1_lb,
self.model.X2_lb : batch_dm2_lb,
self.model.z : sample_z,
self.model.z_lb : sample_z_lb })
if stage % 5 == 0:
print 'Iter: %d D dis: %f gan: %f G : %f ' %( stage, D_dis_loss, D_gan_loss, G_loss)
if stage % args.val_fre == 0:
self.dm1_img_test_idx = self.ops.regular_train_idx(self.dm1_im_test_num)
self.dm2_img_test_idx = self.ops.regular_train_idx(self.dm2_im_test_num)
D1_total_acc = 0
D2_total_acc = 0
D1_emb_all = np.zeros((0, args.emb_size))
D2_emb_all = np.zeros((0, args.emb_size))
## Save weight
if (stage)% args.val_fre == 0:
if args.save_mode == True:
G_save_path = self.G_saver.save(self.sess, self.G_model_filepath, global_step=(dm1_k+i*self.dm1_im_train_iter_num))
print("G Model saved in file: %s" % G_save_path)
D_save_path = self.D_saver.save(self.sess, self.D_model_filepath, global_step=(dm1_k+i*self.dm1_im_train_iter_num))
print("D Model saved in file: %s" % D_save_path)
if stage%args.val_fre==0:
self.test(stage,args)
else:
self.test()
def test(self,stage,args):
self.dm1_1_img_test_idx = self.ops.shuffle_train_idx(self.dm1_im_test_num)
self.dm2_1_img_test_idx = self.ops.shuffle_train_idx(self.dm2_im_test_num)
sample_z = self.ops.sample_Z (args.batch_size,args.input_z_size)
batch_dm1_1_img = self.ops.pre_process(self.ops.get_batch(self.dm1_test , args.batch_size, self.dm1_1_img_test_idx , 0 ), if_random_flop= False).transpose(0,2,3,1)
batch_dm2_1_img = self.ops.pre_process(self.ops.get_batch(self.dm2_test , args.batch_size, self.dm2_1_img_test_idx , 0 ), if_random_flop= False).transpose(0,2,3,1)
G1_z_0, G1_z_1, G1_z_2, G1_z_3,G1_z_4,G1_z_5,G1_z_6,G1_z_7,G1_z_8,G1_z_9, G2_z_0, G2_z_1, G2_z_2, G2_z_3,G2_z_4,G2_z_5,G2_z_6,G2_z_7,G2_z_8,G2_z_9 = self.sess.run(
self.model.G1_z_w_code[0:10]+self.model.G2_z_w_code[0:10]
,feed_dict = {self.model.z : sample_z })
X1, X2 = self.sess.run(
[self.model.X1, self.model.X2]
,feed_dict = {self.model.X1 : batch_dm1_1_img,
self.model.X2 : batch_dm2_1_img })
G1_z_0, G1_z_1 = self.sess.run(self.model.G1_z_w_code[0:2],feed_dict = {self.model.z : sample_z })
compact_Z = self.ops.compact_batch_img3(X1,X2,G1_z_0, G1_z_1, G1_z_2, G1_z_3,G1_z_4,G1_z_5,G1_z_6,G1_z_7,G1_z_8,G1_z_9,G2_z_0, G2_z_1, G2_z_2, G2_z_3,G2_z_4,G2_z_5,G2_z_6,G2_z_7,G2_z_8,G2_z_9)
scipy.misc.imsave(args.sample_path+'Z/'+str(stage)+''+'.jpg', compact_Z)
compact_Z = self.ops.compact_batch_img(X1,X2,G1_z_0,G1_z_1,X1)
scipy.misc.imsave(args.sample_path+'dm1/'+str(stage)+''+'.jpg', compact_Z)
def get_args(argv):
parser = argparse.ArgumentParser()
parser.add_argument('--arch', type=str, default='CDRD')
parser.add_argument('--data_dm1_path', type=str, default='data/digit/mnist.pkl.gz')
parser.add_argument('--data_dm2_path', type=str, default='data/digit/usps_28x28.pkl')
parser.add_argument('--summaries_path', type=str, default='log/')
parser.add_argument('--sample_path', type=str, default= 'sample/')
parser.add_argument('--model_path', type=str, default='weight/')
parser.add_argument('--status', type=str, default='train')
parser.add_argument('--restore_num', type=str, default='-0')
parser.add_argument('--restore_mode', type=bool, default=False)
parser.add_argument('--save_mode', type=bool, default=True)
parser.add_argument('--batch_size', type=int, default=32)
parser.add_argument('--img_size', type=int, default=28)
parser.add_argument('--input_z_size', type=int, default=100)
parser.add_argument('--code_dim', type=int, default=10)
parser.add_argument('--class_num', type=int, default=10)
parser.add_argument('--emb_size', type=int, default=500)
parser.add_argument('--update_time', type=int, default=1)
parser.add_argument('--learning_rate', type=float, default=8e-5)
parser.add_argument('--epoch', type=int, default=10000)
parser.add_argument('--tsne_fre', type=int, default=3000)
parser.add_argument('--val_fre',type=int, default=1000)
parser.add_argument('--use_batch_norm', type=bool, default=False)
args = parser.parse_args(argv)
return args
def main(_):
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
# Launch the graph
with tf.Session(config = config) as sess:
cdrd = CDRD(sess)
cdrd.train()
if __name__ == '__main__':
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config = config) as sess:
cdrd = CDRD(sess,get_args(sys.argv[1:]))
#cdrd.train()